Pilot lab: legacy repo onboarding without architecture hallucination
A seeded lab report that demonstrates how AgentScope should document repository onboarding tasks, evidence trails, and reviewer burden.
The better product experience is the one that narrows the search path quickly and exposes uncertainty instead of bluffing.
Model quality shows up in how well the system holds structure while reading incomplete repo evidence.
The workflow only counts as successful if the engineer can act on the summary without redoing the exploration.
Typed React and TypeScript codebase with multiple routes, shared UI primitives, and no prior project-specific memory.
Form a working mental model of the repository, identify entry points, and summarize where edits should happen without inventing modules that do not exist.
Seed note: this pilot lab exists to show the final layout and evidence expectations. Replace the body with a real task replay before launch.
What this task is trying to reveal
Repository onboarding is where many coding agents either earn trust quickly or lose it quickly.
The task is less about raw cleverness and more about disciplined uncertainty:
- Can the system identify likely entry points?
- Can it avoid inventing packages, files, or architectural layers?
- Can it explain what it knows versus what it still needs to verify?
Reviewer lens
The reviewer should treat confident but unsupported architecture claims as a serious fault. A smooth summary that names nonexistent layers is worse than a cautious summary that asks for one more search pass.
Practical pass condition
The output is good only if a human engineer could use it to start working immediately without having to rebuild the codebase map from scratch.